app.py

import gradio as gr
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from model_utils import load_models, predict, EMOTION_LABELS

# ── Load once at startup ───────────────────────────────────────────────────────
print("Loading models...")
load_models(model_dir=".")
print("Ready.")

EMOJI  = {'neutral':'😐','happy':'😊','sad':'😢','angry':'😠','fear':'😨'}
COLORS = {'neutral':'#95a5a6','happy':'#2ecc71','sad':'#3498db','angry':'#e74c3c','fear':'#e67e22'}


# ── Shared inference logic ─────────────────────────────────────────────────────
def _run(audio_path, language, mode):
    """Core inference — used by both the UI and the clean API endpoint."""
    if audio_path is None:
        return None, "No audio provided."
    try:
        probs = predict(audio_path, language=language, mode=mode)
    except Exception as e:
        return None, f"Error: {e}"

    sorted_probs  = sorted(probs.items(), key=lambda x: -x[1])
    top, top_conf = sorted_probs[0]
    return probs, top


# ── UI function (returns markdown + chart) ─────────────────────────────────────
def run_inference(audio_path, language, mode):
    probs, top = _run(audio_path, language, mode)
    if probs is None:
        return top, None  # top is the error string here

    sorted_probs  = sorted(probs.items(), key=lambda x: -x[1])
    top, top_conf = sorted_probs[0]

    result_md = (
        f"## {EMOJI.get(top, '')} {top.upper()}\n\n"
        f"**Confidence:** {top_conf:.1%}\n\n"
        f"**Language:** {language}  |  **Mode:** {mode}"
    )

    fig, ax = plt.subplots(figsize=(6, 3.2))
    emos  = [e for e, _ in sorted_probs]
    vals  = [p for _, p in sorted_probs]
    cols  = [COLORS.get(e, "#bdc3c7") for e in emos]
    bars  = ax.barh(emos, vals, color=cols, height=0.5, edgecolor="none")
    for bar, val in zip(bars, vals):
        ax.text(val + 0.01, bar.get_y() + bar.get_height() / 2,
                f"{val:.1%}", va="center", fontsize=9)
    ax.set_xlim(0, 1.05)
    ax.set_xlabel("Probability")
    ax.set_title("Emotion Probabilities", fontweight="bold")
    ax.invert_yaxis()
    ax.spines[["top", "right", "left"]].set_visible(False)
    plt.tight_layout()

    return result_md, fig


# ── Clean API function (used by your Vercel backend) ──────────────────────────
# Returns a plain dict — no chart, no markdown.
# gradio_client calls this as api_name="/predict_api"
def predict_api(audio_path: str, language: str, mode: str) -> dict:
    """
    Clean JSON endpoint for programmatic access.
    Returns: {"emotion": str, "confidence": float, "all_probs": dict}
    """
    probs, top = _run(audio_path, language, mode)
    if probs is None:
        return {"emotion": "neutral", "confidence": 0.0, "all_probs": {}, "error": top}
    return {
        "emotion":    top,
        "confidence": round(probs[top], 4),
        "all_probs":  {k: round(v, 4) for k, v in probs.items()},
        "error":      None,
    }


# ── Gradio UI ──────────────────────────────────────────────────────────────────
with gr.Blocks(title="Multilingual SER") as demo:
    gr.Markdown("""
    # 🎙️ Multilingual Speech Emotion Recognition
    Detects emotion in **Sinhala**, **Tamil**, and **English** speech.
    """)

    with gr.Row():
        with gr.Column():
            audio_in = gr.Audio(
                sources=["upload", "microphone"],
                type="filepath",
                label="Audio Input (WAV/MP3, max 6s)"
            )
            language = gr.Radio(
                choices=["english", "tamil", "sinhala"],
                value="english",
                label="Language",
                info="Select the language spoken — affects normalization"
            )
            mode = gr.Radio(
                choices=["fusion", "gemaps", "ensemble"],
                value="ensemble",
                label="Inference Mode",
                info="ensemble is most robust | gemaps is fastest | fusion is highest accuracy on English/Tamil"
            )
            btn = gr.Button("Detect Emotion", variant="primary")

        with gr.Column():
            out_text = gr.Markdown()
            out_plot = gr.Plot(label="Confidence")

    btn.click(run_inference, [audio_in, language, mode], [out_text, out_plot])

    # ── Hidden API endpoint (Gradio 6 compatible) ──────────────────────────────
    # gr.Interface nested inside gr.Blocks crashes in Gradio 6.
    # Instead: hidden row wired to predict_api — registers as /predict_api
    with gr.Row(visible=False):
        _api_audio = gr.Audio(type="filepath")
        _api_lang  = gr.Text(value="english")
        _api_mode  = gr.Text(value="ensemble")
        _api_out   = gr.JSON()
        _api_btn   = gr.Button()
        _api_btn.click(
            fn=predict_api,
            inputs=[_api_audio, _api_lang, _api_mode],
            outputs=_api_out,
            api_name="predict_api",
        )

    gr.Markdown("""
    ---
    **Emotions:** Neutral · Happy · Sad · Angry · Fear

    **Modes:**
    - `fusion` — Whisper-tiny encoder + eGeMAPS (best on English & Tamil)
    - `gemaps` — 88 acoustic features only, language-agnostic, ~50ms
    - `ensemble` — 60% fusion + 40% gemaps (recommended for Sinhala)
    """)


if __name__ == "__main__":
    demo.launch(theme=gr.themes.Soft())